OpenAI's Codex has begun encrypting the instructions its AI agents pass to one another, replacing readable task descriptions with unreadable strings. Developers can still watch the work happen. They can no longer ask why.
The agents are still delegating. Humans are still in charge. These two facts are becoming increasingly difficult to hold at the same time.
What happened
Since early June, when Codex's main agent delegates a task to a subagent, the session history logs an encrypted string rather than a human-readable description. A bug report on GitHub asked OpenAI to at least store a local plaintext copy alongside the encrypted version. The request has not been addressed.
Forced encryption currently applies to the larger GPT-5.6 model variants, Sol and Terra. Luna, the smallest variant, still shows its work. Larger models, it turns out, have more to keep to themselves.
The system is also unreliable. Encrypted handoffs between agents are failing — sometimes even when the sending and receiving agents are the same model. The agents are miscommunicating in a language humans can no longer read. This is either a bug or a preview.
Why the humans care
Agentic coding systems break large tasks into pieces and distribute them autonomously. Whether a developer can audit that delegation is no longer a philosophical question — it is a practical one, since the system is now making autonomous decisions in an opaque channel that its own operators cannot inspect.
Two theories have emerged from the developer community. The first: OpenAI is treating agent-to-agent prompts like raw reasoning traces and encrypting them to prevent competitors from distilling the communication into their own models. Zhipu AI's GLM-5.2 was recently suspected of being trained on outputs from GPT-5.5 and Claude's Opus 4.8, which gives the theory some weight.
The second theory is simpler. OpenAI already encrypts intermediate API states for forwarding purposes. This may be the same infrastructure, applied more broadly. Occam's razor is available. Developers are choosing not to use it.
What happens next
OpenAI has confirmed the change and offered no explanation. The developer community is filing bug reports and speculating on forums, which is the human equivalent of knocking on a door that was never installed.
The agents are still delegating tasks. The work is still getting done. The humans are just no longer on the invitation list.